"vscode:/vscode.git/clone" did not exist on "51cab350d27185ac1286dddd1fdee87994dfc66d"
scheduling_dpmsolver_singlestep.py 53.6 KB
Newer Older
Aryan's avatar
Aryan committed
1
# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver

import math
18
from typing import List, Literal, Optional, Tuple, Union
19
20
21
22
23

import numpy as np
import torch

from ..configuration_utils import ConfigMixin, register_to_config
24
from ..utils import deprecate, is_scipy_available, logging
25
from ..utils.torch_utils import randn_tensor
Kashif Rasul's avatar
Kashif Rasul committed
26
from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
27
28


29
30
31
if is_scipy_available():
    import scipy.stats

Patrick von Platen's avatar
Patrick von Platen committed
32
33
34
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


35
# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
YiYi Xu's avatar
YiYi Xu committed
36
def betas_for_alpha_bar(
37
38
39
40
    num_diffusion_timesteps: int,
    max_beta: float = 0.999,
    alpha_transform_type: Literal["cosine", "exp"] = "cosine",
) -> torch.Tensor:
41
42
43
44
45
46
47
48
    """
    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
    (1-beta) over time from t = [0,1].

    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
    to that part of the diffusion process.

    Args:
49
50
51
52
53
54
        num_diffusion_timesteps (`int`):
            The number of betas to produce.
        max_beta (`float`, defaults to `0.999`):
            The maximum beta to use; use values lower than 1 to avoid numerical instability.
        alpha_transform_type (`"cosine"` or `"exp"`, defaults to `"cosine"`):
            The type of noise schedule for `alpha_bar`. Choose from `cosine` or `exp`.
55
56

    Returns:
57
58
        `torch.Tensor`:
            The betas used by the scheduler to step the model outputs.
59
    """
YiYi Xu's avatar
YiYi Xu committed
60
    if alpha_transform_type == "cosine":
61

YiYi Xu's avatar
YiYi Xu committed
62
63
64
65
66
67
68
69
70
        def alpha_bar_fn(t):
            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2

    elif alpha_transform_type == "exp":

        def alpha_bar_fn(t):
            return math.exp(t * -12.0)

    else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
71
        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
72
73
74
75
76

    betas = []
    for i in range(num_diffusion_timesteps):
        t1 = i / num_diffusion_timesteps
        t2 = (i + 1) / num_diffusion_timesteps
YiYi Xu's avatar
YiYi Xu committed
77
        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
78
79
80
81
82
    return torch.tensor(betas, dtype=torch.float32)


class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
    """
83
    `DPMSolverSinglestepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
84

85
86
    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
    methods the library implements for all schedulers such as loading and saving.
87
88

    Args:
89
90
91
92
93
94
95
96
        num_train_timesteps (`int`, defaults to 1000):
            The number of diffusion steps to train the model.
        beta_start (`float`, defaults to 0.0001):
            The starting `beta` value of inference.
        beta_end (`float`, defaults to 0.02):
            The final `beta` value.
        beta_schedule (`str`, defaults to `"linear"`):
            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
97
            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
98
99
100
101
        trained_betas (`np.ndarray`, *optional*):
            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
        solver_order (`int`, defaults to 2):
            The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
102
            sampling, and `solver_order=3` for unconditional sampling.
103
104
105
106
107
108
109
110
111
112
113
114
115
        prediction_type (`str`, defaults to `epsilon`, *optional*):
            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
            Video](https://imagen.research.google/video/paper.pdf) paper).
        thresholding (`bool`, defaults to `False`):
            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
            as Stable Diffusion.
        dynamic_thresholding_ratio (`float`, defaults to 0.995):
            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
        sample_max_value (`float`, defaults to 1.0):
            The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
            `algorithm_type="dpmsolver++"`.
        algorithm_type (`str`, defaults to `dpmsolver++`):
116
117
118
119
120
            Algorithm type for the solver; can be `dpmsolver` or `dpmsolver++` or `sde-dpmsolver++`. The `dpmsolver`
            type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927) paper, and the
            `dpmsolver++` type implements the algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095)
            paper. It is recommended to use `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided
            sampling like in Stable Diffusion.
121
122
123
124
125
126
        solver_type (`str`, defaults to `midpoint`):
            Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
            sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
        lower_order_final (`bool`, defaults to `True`):
            Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
            stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
127
        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
128
129
            Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
            the sigmas are determined according to a sequence of noise levels {σi}.
130
131
        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
132
133
134
        use_beta_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta
            Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information.
135
        final_sigmas_type (`str`, *optional*, defaults to `"zero"`):
136
137
            The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final
            sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
138
139
140
        lambda_min_clipped (`float`, defaults to `-inf`):
            Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
            cosine (`squaredcos_cap_v2`) noise schedule.
141
        variance_type (`str`, *optional*):
142
143
            Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
            contains the predicted Gaussian variance.
144
145
    """

Kashif Rasul's avatar
Kashif Rasul committed
146
    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
    order = 1

    @register_to_config
    def __init__(
        self,
        num_train_timesteps: int = 1000,
        beta_start: float = 0.0001,
        beta_end: float = 0.02,
        beta_schedule: str = "linear",
        trained_betas: Optional[np.ndarray] = None,
        solver_order: int = 2,
        prediction_type: str = "epsilon",
        thresholding: bool = False,
        dynamic_thresholding_ratio: float = 0.995,
        sample_max_value: float = 1.0,
        algorithm_type: str = "dpmsolver++",
        solver_type: str = "midpoint",
164
        lower_order_final: bool = False,
165
        use_karras_sigmas: Optional[bool] = False,
166
        use_exponential_sigmas: Optional[bool] = False,
167
        use_beta_sigmas: Optional[bool] = False,
168
169
        use_flow_sigmas: Optional[bool] = False,
        flow_shift: Optional[float] = 1.0,
170
        final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
171
172
        lambda_min_clipped: float = -float("inf"),
        variance_type: Optional[str] = None,
173
174
        use_dynamic_shifting: bool = False,
        time_shift_type: str = "exponential",
175
    ):
176
177
178
179
180
181
        if self.config.use_beta_sigmas and not is_scipy_available():
            raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
        if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
            raise ValueError(
                "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used."
            )
182
183
184
185
        if algorithm_type == "dpmsolver":
            deprecation_message = "algorithm_type `dpmsolver` is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
            deprecate("algorithm_types=dpmsolver", "1.0.0", deprecation_message)

186
187
188
189
190
191
        if trained_betas is not None:
            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
        elif beta_schedule == "linear":
            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
        elif beta_schedule == "scaled_linear":
            # this schedule is very specific to the latent diffusion model.
192
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
193
194
195
196
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
197
            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
198
199
200
201
202
203
204

        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
        # Currently we only support VP-type noise schedule
        self.alpha_t = torch.sqrt(self.alphas_cumprod)
        self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
        self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
205
        self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
206
207
208
209
210

        # standard deviation of the initial noise distribution
        self.init_noise_sigma = 1.0

        # settings for DPM-Solver
211
        if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver++"]:
212
            if algorithm_type == "deis":
213
                self.register_to_config(algorithm_type="dpmsolver++")
214
            else:
215
                raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
216
        if solver_type not in ["midpoint", "heun"]:
217
            if solver_type in ["logrho", "bh1", "bh2"]:
218
                self.register_to_config(solver_type="midpoint")
219
            else:
220
                raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}")
221

222
        if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero":
223
            raise ValueError(
224
                f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead."
225
226
            )

227
228
229
230
231
232
233
        # setable values
        self.num_inference_steps = None
        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
        self.timesteps = torch.from_numpy(timesteps)
        self.model_outputs = [None] * solver_order
        self.sample = None
        self.order_list = self.get_order_list(num_train_timesteps)
234
        self._step_index = None
235
        self._begin_index = None
236
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
237
238
239
240
241
242
243

    def get_order_list(self, num_inference_steps: int) -> List[int]:
        """
        Computes the solver order at each time step.

        Args:
            num_inference_steps (`int`):
244
                The number of diffusion steps used when generating samples with a pre-trained model.
245
246
        """
        steps = num_inference_steps
247
        order = self.config.solver_order
248
249
        if order > 3:
            raise ValueError("Order > 3 is not supported by this scheduler")
250
        if self.config.lower_order_final:
251
252
253
254
255
256
257
258
259
            if order == 3:
                if steps % 3 == 0:
                    orders = [1, 2, 3] * (steps // 3 - 1) + [1, 2] + [1]
                elif steps % 3 == 1:
                    orders = [1, 2, 3] * (steps // 3) + [1]
                else:
                    orders = [1, 2, 3] * (steps // 3) + [1, 2]
            elif order == 2:
                if steps % 2 == 0:
260
                    orders = [1, 2] * (steps // 2 - 1) + [1, 1]
261
262
263
264
265
266
267
268
269
270
271
                else:
                    orders = [1, 2] * (steps // 2) + [1]
            elif order == 1:
                orders = [1] * steps
        else:
            if order == 3:
                orders = [1, 2, 3] * (steps // 3)
            elif order == 2:
                orders = [1, 2] * (steps // 2)
            elif order == 1:
                orders = [1] * steps
StAlKeR7779's avatar
StAlKeR7779 committed
272
273
274
275

        if self.config.final_sigmas_type == "zero":
            orders[-1] = 1

276
277
        return orders

278
279
280
    @property
    def step_index(self):
        """
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
281
        The index counter for current timestep. It will increase 1 after each scheduler step.
282
283
284
        """
        return self._step_index

285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
    @property
    def begin_index(self):
        """
        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
        """
        return self._begin_index

    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
    def set_begin_index(self, begin_index: int = 0):
        """
        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.

        Args:
            begin_index (`int`):
                The begin index for the scheduler.
        """
        self._begin_index = begin_index

303
304
305
306
    def set_timesteps(
        self,
        num_inference_steps: int = None,
        device: Union[str, torch.device] = None,
307
        mu: Optional[float] = None,
308
309
        timesteps: Optional[List[int]] = None,
    ):
310
        """
311
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
312
313
314

        Args:
            num_inference_steps (`int`):
315
316
317
                The number of diffusion steps used when generating samples with a pre-trained model.
            device (`str` or `torch.device`, *optional*):
                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
318
319
320
321
            timesteps (`List[int]`, *optional*):
                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
                timestep spacing strategy of equal spacing between timesteps schedule is used. If `timesteps` is
                passed, `num_inference_steps` must be `None`.
322
        """
323
324
325
        if mu is not None:
            assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential"
            self.config.flow_shift = np.exp(mu)
326
327
328
329
330
331
        if num_inference_steps is None and timesteps is None:
            raise ValueError("Must pass exactly one of  `num_inference_steps` or `timesteps`.")
        if num_inference_steps is not None and timesteps is not None:
            raise ValueError("Must pass exactly one of  `num_inference_steps` or `timesteps`.")
        if timesteps is not None and self.config.use_karras_sigmas:
            raise ValueError("Cannot use `timesteps` when `config.use_karras_sigmas=True`.")
332
333
        if timesteps is not None and self.config.use_exponential_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.")
334
335
        if timesteps is not None and self.config.use_beta_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.")
336
337

        num_inference_steps = num_inference_steps or len(timesteps)
338
        self.num_inference_steps = num_inference_steps
339
340
341
342
343
344
345

        if timesteps is not None:
            timesteps = np.array(timesteps).astype(np.int64)
        else:
            # Clipping the minimum of all lambda(t) for numerical stability.
            # This is critical for cosine (squaredcos_cap_v2) noise schedule.
            clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped)
YiYi Xu's avatar
YiYi Xu committed
346
            clipped_idx = clipped_idx.item()
347
348
349
350
351
352
            timesteps = (
                np.linspace(0, self.config.num_train_timesteps - 1 - clipped_idx, num_inference_steps + 1)
                .round()[::-1][:-1]
                .copy()
                .astype(np.int64)
            )
353

354
        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
355
        log_sigmas = np.log(sigmas)
356
        if self.config.use_karras_sigmas:
357
            sigmas = np.flip(sigmas).copy()
358
359
            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
360
        elif self.config.use_exponential_sigmas:
361
362
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
363
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
364
        elif self.config.use_beta_sigmas:
365
366
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
367
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
368
369
370
        elif self.config.use_flow_sigmas:
            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
            sigmas = 1.0 - alphas
hlky's avatar
hlky committed
371
            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy()
372
            timesteps = (sigmas * self.config.num_train_timesteps).copy()
373
374
        else:
            sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
375
376

        if self.config.final_sigmas_type == "sigma_min":
377
            sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
378
379
380
381
382
383
384
        elif self.config.final_sigmas_type == "zero":
            sigma_last = 0
        else:
            raise ValueError(
                f" `final_sigmas_type` must be one of `sigma_min` or `zero`, but got {self.config.final_sigmas_type}"
            )
        sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
385

386
        self.sigmas = torch.from_numpy(sigmas).to(device=device)
387

388
        self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
389
390
        self.model_outputs = [None] * self.config.solver_order
        self.sample = None
Patrick von Platen's avatar
Patrick von Platen committed
391
392

        if not self.config.lower_order_final and num_inference_steps % self.config.solver_order != 0:
393
            logger.warning(
394
                "Changing scheduler {self.config} to have `lower_order_final` set to True to handle uneven amount of inference steps. Please make sure to always use an even number of `num_inference steps when using `lower_order_final=False`."
Patrick von Platen's avatar
Patrick von Platen committed
395
396
397
            )
            self.register_to_config(lower_order_final=True)

398
        if not self.config.lower_order_final and self.config.final_sigmas_type == "zero":
399
            logger.warning(
400
401
402
403
                " `last_sigmas_type='zero'` is not supported for `lower_order_final=False`. Changing scheduler {self.config} to have `lower_order_final` set to True."
            )
            self.register_to_config(lower_order_final=True)

Patrick von Platen's avatar
Patrick von Platen committed
404
        self.order_list = self.get_order_list(num_inference_steps)
405

406
407
        # add an index counter for schedulers that allow duplicated timesteps
        self._step_index = None
408
        self._begin_index = None
409
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
410

411
    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
412
    def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor:
413
        """
414
415
        Apply dynamic thresholding to the predicted sample.

416
417
418
419
420
421
        "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
        prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
        s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
        pixels from saturation at each step. We find that dynamic thresholding results in significantly better
        photorealism as well as better image-text alignment, especially when using very large guidance weights."

Quentin Gallouédec's avatar
Quentin Gallouédec committed
422
        https://huggingface.co/papers/2205.11487
423
424
425
426
427
428
429
430

        Args:
            sample (`torch.Tensor`):
                The predicted sample to be thresholded.

        Returns:
            `torch.Tensor`:
                The thresholded sample.
431
432
        """
        dtype = sample.dtype
433
        batch_size, channels, *remaining_dims = sample.shape
434
435
436
437
438

        if dtype not in (torch.float32, torch.float64):
            sample = sample.float()  # upcast for quantile calculation, and clamp not implemented for cpu half

        # Flatten sample for doing quantile calculation along each image
439
        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
440
441
442
443
444
445
446
447
448
449

        abs_sample = sample.abs()  # "a certain percentile absolute pixel value"

        s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
        s = torch.clamp(
            s, min=1, max=self.config.sample_max_value
        )  # When clamped to min=1, equivalent to standard clipping to [-1, 1]
        s = s.unsqueeze(1)  # (batch_size, 1) because clamp will broadcast along dim=0
        sample = torch.clamp(sample, -s, s) / s  # "we threshold xt0 to the range [-s, s] and then divide by s"

450
        sample = sample.reshape(batch_size, channels, *remaining_dims)
451
452
453
        sample = sample.to(dtype)

        return sample
454

455
456
457
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
    def _sigma_to_t(self, sigma, log_sigmas):
        # get log sigma
458
        log_sigma = np.log(np.maximum(sigma, 1e-10))
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478

        # get distribution
        dists = log_sigma - log_sigmas[:, np.newaxis]

        # get sigmas range
        low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
        high_idx = low_idx + 1

        low = log_sigmas[low_idx]
        high = log_sigmas[high_idx]

        # interpolate sigmas
        w = (low - log_sigma) / (low - high)
        w = np.clip(w, 0, 1)

        # transform interpolation to time range
        t = (1 - w) * low_idx + w * high_idx
        t = t.reshape(sigma.shape)
        return t

479
480
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
    def _sigma_to_alpha_sigma_t(self, sigma):
481
482
483
484
485
486
        if self.config.use_flow_sigmas:
            alpha_t = 1 - sigma
            sigma_t = sigma
        else:
            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
            sigma_t = sigma * alpha_t
487
488
489

        return alpha_t, sigma_t

490
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
491
    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
492
493
        """Constructs the noise schedule of Karras et al. (2022)."""

Suraj Patil's avatar
Suraj Patil committed
494
495
496
497
498
499
500
501
502
503
504
505
506
507
        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
508
509
510
511
512
513
514
515

        rho = 7.0  # 7.0 is the value used in the paper
        ramp = np.linspace(0, 1, num_inference_steps)
        min_inv_rho = sigma_min ** (1 / rho)
        max_inv_rho = sigma_max ** (1 / rho)
        sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
        return sigmas

516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential
    def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
        """Constructs an exponential noise schedule."""

        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()

535
        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
536
537
        return sigmas

538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta
    def _convert_to_beta(
        self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6
    ) -> torch.Tensor:
        """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)"""

        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()

559
        sigmas = np.array(
560
561
562
563
564
565
566
567
568
569
            [
                sigma_min + (ppf * (sigma_max - sigma_min))
                for ppf in [
                    scipy.stats.beta.ppf(timestep, alpha, beta)
                    for timestep in 1 - np.linspace(0, 1, num_inference_steps)
                ]
            ]
        )
        return sigmas

570
    def convert_model_output(
571
        self,
572
        model_output: torch.Tensor,
573
        *args,
574
        sample: torch.Tensor = None,
575
        **kwargs,
576
    ) -> torch.Tensor:
577
        """
578
579
580
581
        Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
        designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
        integral of the data prediction model.

Steven Liu's avatar
Steven Liu committed
582
583
        > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both
        noise > prediction and data prediction models.
584
585

        Args:
586
            model_output (`torch.Tensor`):
587
                The direct output from the learned diffusion model.
588
            sample (`torch.Tensor`):
589
                A current instance of a sample created by the diffusion process.
590
591

        Returns:
592
            `torch.Tensor`:
593
                The converted model output.
594
        """
595
596
597
598
599
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        if sample is None:
            if len(args) > 1:
                sample = args[1]
            else:
600
                raise ValueError("missing `sample` as a required keyword argument")
601
602
603
604
605
606
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
607
        # DPM-Solver++ needs to solve an integral of the data prediction model.
608
        if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
609
            if self.config.prediction_type == "epsilon":
610
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
611
                if self.config.variance_type in ["learned", "learned_range"]:
612
                    model_output = model_output[:, :3]
613
614
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
615
616
617
618
                x0_pred = (sample - sigma_t * model_output) / alpha_t
            elif self.config.prediction_type == "sample":
                x0_pred = model_output
            elif self.config.prediction_type == "v_prediction":
619
620
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
621
                x0_pred = alpha_t * sample - sigma_t * model_output
622
623
624
            elif self.config.prediction_type == "flow_prediction":
                sigma_t = self.sigmas[self.step_index]
                x0_pred = sample - sigma_t * model_output
625
626
            else:
                raise ValueError(
627
628
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
                    "`v_prediction`, or `flow_prediction` for the DPMSolverSinglestepScheduler."
629
630
631
                )

            if self.config.thresholding:
632
633
                x0_pred = self._threshold_sample(x0_pred)

634
            return x0_pred
635

636
637
638
        # DPM-Solver needs to solve an integral of the noise prediction model.
        elif self.config.algorithm_type == "dpmsolver":
            if self.config.prediction_type == "epsilon":
639
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
640
641
642
643
                if self.config.variance_type in ["learned", "learned_range"]:
                    epsilon = model_output[:, :3]
                else:
                    epsilon = model_output
644
            elif self.config.prediction_type == "sample":
645
646
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
647
648
                epsilon = (sample - alpha_t * model_output) / sigma_t
            elif self.config.prediction_type == "v_prediction":
649
650
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
651
652
653
654
655
656
657
                epsilon = alpha_t * model_output + sigma_t * sample
            else:
                raise ValueError(
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
                    " `v_prediction` for the DPMSolverSinglestepScheduler."
                )

658
659
660
661
662
663
664
665
            if self.config.thresholding:
                alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep]
                x0_pred = (sample - sigma_t * epsilon) / alpha_t
                x0_pred = self._threshold_sample(x0_pred)
                epsilon = (sample - alpha_t * x0_pred) / sigma_t

            return epsilon

666
667
    def dpm_solver_first_order_update(
        self,
668
        model_output: torch.Tensor,
669
        *args,
670
        sample: torch.Tensor = None,
671
        noise: Optional[torch.Tensor] = None,
672
        **kwargs,
673
    ) -> torch.Tensor:
674
        """
675
        One step for the first-order DPMSolver (equivalent to DDIM).
676
677

        Args:
678
            model_output (`torch.Tensor`):
679
680
681
682
683
                The direct output from the learned diffusion model.
            timestep (`int`):
                The current discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
684
            sample (`torch.Tensor`):
685
                A current instance of a sample created by the diffusion process.
686
687

        Returns:
688
            `torch.Tensor`:
689
                The sample tensor at the previous timestep.
690
        """
691
692
693
694
695
696
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
697
                raise ValueError("missing `sample` as a required keyword argument")
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
        sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
716
717
718
719
720
        h = lambda_t - lambda_s
        if self.config.algorithm_type == "dpmsolver++":
            x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
        elif self.config.algorithm_type == "dpmsolver":
            x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
721
722
723
724
725
726
727
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            x_t = (
                (sigma_t / sigma_s * torch.exp(-h)) * sample
                + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
            )
728
729
730
731
        return x_t

    def singlestep_dpm_solver_second_order_update(
        self,
732
        model_output_list: List[torch.Tensor],
733
        *args,
734
        sample: torch.Tensor = None,
735
        noise: Optional[torch.Tensor] = None,
736
        **kwargs,
737
    ) -> torch.Tensor:
738
        """
739
740
        One step for the second-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the
        time `timestep_list[-2]`.
741
742

        Args:
743
            model_output_list (`List[torch.Tensor]`):
744
745
746
747
748
                The direct outputs from learned diffusion model at current and latter timesteps.
            timestep (`int`):
                The current and latter discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
749
            sample (`torch.Tensor`):
750
                A current instance of a sample created by the diffusion process.
751
752

        Returns:
753
            `torch.Tensor`:
754
                The sample tensor at the previous timestep.
755
        """
756
757
758
759
760
761
        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
762
                raise ValueError("missing `sample` as a required keyword argument")
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
        sigma_t, sigma_s0, sigma_s1 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
        )

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)

790
        m0, m1 = model_output_list[-1], model_output_list[-2]
791

792
793
794
795
        h, h_0 = lambda_t - lambda_s1, lambda_s0 - lambda_s1
        r0 = h_0 / h
        D0, D1 = m1, (1.0 / r0) * (m0 - m1)
        if self.config.algorithm_type == "dpmsolver++":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
796
            # See https://huggingface.co/papers/2211.01095 for detailed derivations
797
798
799
800
801
802
803
804
805
806
807
808
809
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s1) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s1) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                )
        elif self.config.algorithm_type == "dpmsolver":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
810
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
811
812
813
814
815
816
817
818
819
820
821
822
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s1) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s1) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                )
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s1 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s1 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
839
840
841
842
        return x_t

    def singlestep_dpm_solver_third_order_update(
        self,
843
        model_output_list: List[torch.Tensor],
844
        *args,
845
        sample: torch.Tensor = None,
StAlKeR7779's avatar
StAlKeR7779 committed
846
        noise: Optional[torch.Tensor] = None,
847
        **kwargs,
848
    ) -> torch.Tensor:
849
        """
850
851
        One step for the third-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the
        time `timestep_list[-3]`.
852
853

        Args:
854
            model_output_list (`List[torch.Tensor]`):
855
856
857
858
859
                The direct outputs from learned diffusion model at current and latter timesteps.
            timestep (`int`):
                The current and latter discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
860
            sample (`torch.Tensor`):
861
                A current instance of a sample created by diffusion process.
862
863

        Returns:
864
            `torch.Tensor`:
865
                The sample tensor at the previous timestep.
866
        """
867
868
869
870
871
872
873

        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
874
                raise ValueError("missing `sample` as a required keyword argument")
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
            self.sigmas[self.step_index - 2],
894
        )
895
896
897
898
899
900
901
902
903
904
905
906
907

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
        alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
        lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)

        m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]

908
909
910
911
912
913
914
        h, h_0, h_1 = lambda_t - lambda_s2, lambda_s0 - lambda_s2, lambda_s1 - lambda_s2
        r0, r1 = h_0 / h, h_1 / h
        D0 = m2
        D1_0, D1_1 = (1.0 / r1) * (m1 - m2), (1.0 / r0) * (m0 - m2)
        D1 = (r0 * D1_0 - r1 * D1_1) / (r0 - r1)
        D2 = 2.0 * (D1_1 - D1_0) / (r0 - r1)
        if self.config.algorithm_type == "dpmsolver++":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
915
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
916
917
918
919
920
921
922
923
924
925
926
927
928
929
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s2) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1_1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s2) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                    - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
                )
        elif self.config.algorithm_type == "dpmsolver":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
930
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
931
932
933
934
935
936
937
938
939
940
941
942
943
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s2) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1_1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s2) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                    - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
                )
StAlKeR7779's avatar
StAlKeR7779 committed
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s2 * torch.exp(-h)) * sample
                    + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1_1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s2 * torch.exp(-h)) * sample
                    + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h) + (-2.0 * h)) / (-2.0 * h) ** 2 - 0.5)) * D2
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
961
962
963
964
        return x_t

    def singlestep_dpm_solver_update(
        self,
965
        model_output_list: List[torch.Tensor],
966
        *args,
967
        sample: torch.Tensor = None,
968
        order: int = None,
969
        noise: Optional[torch.Tensor] = None,
970
        **kwargs,
971
    ) -> torch.Tensor:
972
        """
973
        One step for the singlestep DPMSolver.
974
975

        Args:
976
            model_output_list (`List[torch.Tensor]`):
977
978
979
980
981
                The direct outputs from learned diffusion model at current and latter timesteps.
            timestep (`int`):
                The current and latter discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
982
            sample (`torch.Tensor`):
983
                A current instance of a sample created by diffusion process.
984
            order (`int`):
985
                The solver order at this step.
986
987

        Returns:
988
            `torch.Tensor`:
989
                The sample tensor at the previous timestep.
990
        """
991
992
993
994
995
996
        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
997
                raise ValueError("missing `sample` as a required keyword argument")
998
999
1000
1001
        if order is None:
            if len(args) > 3:
                order = args[3]
            else:
1002
                raise ValueError("missing `order` as a required keyword argument")
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

1017
        if order == 1:
1018
            return self.dpm_solver_first_order_update(model_output_list[-1], sample=sample, noise=noise)
1019
        elif order == 2:
1020
            return self.singlestep_dpm_solver_second_order_update(model_output_list, sample=sample, noise=noise)
1021
        elif order == 3:
StAlKeR7779's avatar
StAlKeR7779 committed
1022
            return self.singlestep_dpm_solver_third_order_update(model_output_list, sample=sample, noise=noise)
1023
1024
1025
        else:
            raise ValueError(f"Order must be 1, 2, 3, got {order}")

1026
1027
1028
1029
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep
    def index_for_timestep(self, timestep, schedule_timesteps=None):
        if schedule_timesteps is None:
            schedule_timesteps = self.timesteps
1030

1031
        index_candidates = (schedule_timesteps == timestep).nonzero()
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043

        if len(index_candidates) == 0:
            step_index = len(self.timesteps) - 1
        # The sigma index that is taken for the **very** first `step`
        # is always the second index (or the last index if there is only 1)
        # This way we can ensure we don't accidentally skip a sigma in
        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
        elif len(index_candidates) > 1:
            step_index = index_candidates[1].item()
        else:
            step_index = index_candidates[0].item()

1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
        return step_index

    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
    def _init_step_index(self, timestep):
        """
        Initialize the step_index counter for the scheduler.
        """

        if self.begin_index is None:
            if isinstance(timestep, torch.Tensor):
                timestep = timestep.to(self.timesteps.device)
            self._step_index = self.index_for_timestep(timestep)
        else:
            self._step_index = self._begin_index
1058

1059
1060
    def step(
        self,
1061
        model_output: torch.Tensor,
1062
        timestep: Union[int, torch.Tensor],
1063
        sample: torch.Tensor,
1064
        generator=None,
1065
1066
1067
        return_dict: bool = True,
    ) -> Union[SchedulerOutput, Tuple]:
        """
1068
1069
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
        the singlestep DPMSolver.
1070
1071

        Args:
1072
            model_output (`torch.Tensor`):
1073
1074
1075
                The direct output from learned diffusion model.
            timestep (`int`):
                The current discrete timestep in the diffusion chain.
1076
            sample (`torch.Tensor`):
1077
1078
1079
                A current instance of a sample created by the diffusion process.
            return_dict (`bool`):
                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
1080
1081

        Returns:
1082
1083
1084
            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
                If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
                tuple is returned where the first element is the sample tensor.
1085
1086
1087
1088
1089
1090
1091

        """
        if self.num_inference_steps is None:
            raise ValueError(
                "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
            )

1092
1093
        if self.step_index is None:
            self._init_step_index(timestep)
1094

1095
        model_output = self.convert_model_output(model_output, sample=sample)
1096
1097
1098
1099
        for i in range(self.config.solver_order - 1):
            self.model_outputs[i] = self.model_outputs[i + 1]
        self.model_outputs[-1] = model_output

1100
1101
1102
1103
1104
1105
1106
        if self.config.algorithm_type == "sde-dpmsolver++":
            noise = randn_tensor(
                model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
            )
        else:
            noise = None

1107
        order = self.order_list[self.step_index]
Patrick von Platen's avatar
Patrick von Platen committed
1108
1109
1110
1111
1112
1113

        #  For img2img denoising might start with order>1 which is not possible
        #  In this case make sure that the first two steps are both order=1
        while self.model_outputs[-order] is None:
            order -= 1

1114
1115
1116
1117
        # For single-step solvers, we use the initial value at each time with order = 1.
        if order == 1:
            self.sample = sample

1118
1119
1120
        prev_sample = self.singlestep_dpm_solver_update(
            self.model_outputs, sample=self.sample, order=order, noise=noise
        )
1121

1122
        # upon completion increase step index by one, noise=noise
1123
        self._step_index += 1
1124
1125
1126
1127
1128
1129

        if not return_dict:
            return (prev_sample,)

        return SchedulerOutput(prev_sample=prev_sample)

1130
    def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor:
1131
1132
1133
1134
1135
        """
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep.

        Args:
1136
            sample (`torch.Tensor`):
1137
                The input sample.
1138
1139

        Returns:
1140
            `torch.Tensor`:
1141
                A scaled input sample.
1142
1143
1144
        """
        return sample

1145
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise
1146
1147
    def add_noise(
        self,
1148
1149
        original_samples: torch.Tensor,
        noise: torch.Tensor,
1150
        timesteps: torch.IntTensor,
1151
    ) -> torch.Tensor:
1152
1153
1154
1155
1156
1157
1158
1159
1160
        # Make sure sigmas and timesteps have the same device and dtype as original_samples
        sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
            # mps does not support float64
            schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
        else:
            schedule_timesteps = self.timesteps.to(original_samples.device)
            timesteps = timesteps.to(original_samples.device)
1161

1162
        # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index
1163
1164
        if self.begin_index is None:
            step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
1165
1166
1167
        elif self.step_index is not None:
            # add_noise is called after first denoising step (for inpainting)
            step_indices = [self.step_index] * timesteps.shape[0]
1168
        else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
1169
            # add noise is called before first denoising step to create initial latent(img2img)
1170
            step_indices = [self.begin_index] * timesteps.shape[0]
1171

1172
1173
1174
        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < len(original_samples.shape):
            sigma = sigma.unsqueeze(-1)
1175

1176
1177
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
        noisy_samples = alpha_t * original_samples + sigma_t * noise
1178
1179
1180
1181
        return noisy_samples

    def __len__(self):
        return self.config.num_train_timesteps